File size: 108,944 Bytes
d5eed08
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
"""
2025.3.12
2025.3.14
4.48.3
0.15.2
__UNSLOTH_VERSIONING__
"""
from torch import Tensor
import torch
import torch.nn as nn
from torch.nn import functional as F
from trl.trainer.dpo_trainer import (Any, AutoModelForCausalLM, BaseImageProcessor, Callable, DPOConfig, DPOTrainer, DataCollator, DataCollatorForPreference, DataLoader, Dataset, EvalLoopOutput, F, FDivergenceConstants, FDivergenceType, FeatureExtractionMixin, IterableDataset, Literal, MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES, Optional, PartialState, PeftModel, PreTrainedModel, PreTrainedModelWrapper, PreTrainedTokenizerBase, ProcessorMixin, RunningMoments, SyncRefModelCallback, Trainer, TrainerCallback, Union, amp, cap_exp, contextmanager, create_reference_model, dataclass, deepcopy, defaultdict, deprecate_kwarg, disable_dropout_in_model, empty_cache, flush_left, generate_model_card, get_comet_experiment_url, inspect, is_comet_available, is_peft_available, is_torch_xpu_available, is_wandb_available, log_table_to_comet_experiment, maybe_apply_chat_template, maybe_extract_prompt, nn, nullcontext, os, pad, pad_to_length, pd, peft_module_casting_to_bf16, prepare_model_for_kbit_training, random, textwrap, torch, tqdm, transformers, version, wandb, warnings)


import os
from typing import *
from dataclasses import dataclass, field
from packaging.version import Version
import torch
import numpy as np
from contextlib import nullcontext
from torch.nn import functional as F
from transformers import DataCollatorForSeq2Seq, DataCollatorForLanguageModeling

torch_compile_options = {
    "epilogue_fusion"   : True,
    "max_autotune"      : False,
    "shape_padding"     : True,
    "trace.enabled"     : False,
    "triton.cudagraphs" : False,
}

@torch.compile(dynamic = True, fullgraph = True, options = torch_compile_options,)
def selective_log_softmax(logits, index):
    logits = logits.to(torch.float32)
    selected_logits = torch.gather(logits, dim = -1, index = index.unsqueeze(-1)).squeeze(-1)
    # loop to reduce peak mem consumption
    # logsumexp_values = torch.stack([torch.logsumexp(lg, dim=-1) for lg in logits])
    logsumexp_values = torch.logsumexp(logits, dim = -1)
    per_token_logps = selected_logits - logsumexp_values  # log_softmax(x_i) = x_i - logsumexp(x)
    return per_token_logps
@dataclass
class UnslothDPOConfig(DPOConfig):
    """
    
    Configuration class for the [`DPOTrainer`].

    Using [`~transformers.HfArgumentParser`] we can turn this class into
    [argparse](https://docs.python.org/3/library/argparse#module-argparse) arguments that can be specified on the
    command line.

    Parameters:
        > Parameters that control the model and reference model

        model_init_kwargs (`dict[str, Any]` or `None`, *optional*, defaults to `None`):
            Keyword arguments for `AutoModelForCausalLM.from_pretrained`, used when the `model` argument of the
            [`DPOTrainer`] is provided as a string.
        ref_model_init_kwargs (`dict[str, Any]` or `None`, *optional*, defaults to `None`):
            Keyword arguments for `AutoModelForCausalLM.from_pretrained`, used when the `ref_model` argument of the
            [`DPOTrainer`] is provided as a string.
        model_adapter_name (`str` or `None`, *optional*, defaults to `None`):
            Name of the train target PEFT adapter, when using LoRA with multiple adapters.
        ref_adapter_name (`str` or `None`, *optional*, defaults to `None`):
            Name of the reference PEFT adapter, when using LoRA with multiple adapters.
        force_use_ref_model (`bool`, *optional*, defaults to `False`):
            If you provide a PEFT model as the active model and wish to use a different model for the `ref_model`, set
            this flag to `True`.
        disable_dropout (`bool`, *optional*, defaults to `True`):
            Whether to disable dropout in the model and reference model.
        use_logits_to_keep (`bool`, *optional*, defaults to `False`):
            If `True`, only a specified number of logits are computed in the forward pass. This can be useful for
            saving memory and speeding up training by not computing the logits for all tokens, especially in
            scenarios when working with very long prompts where labels are ignored (-100).

        > Parameters that control the data preprocessing

        dataset_num_proc (`int` or `None`, *optional*, defaults to `None`):
            Number of processes to use for processing the dataset.
        padding_value (`int` or `None`, *optional*, defaults to `None`):
            Padding value to use. If `None`, the padding value of the tokenizer is used.
        label_pad_token_id (`int`, *optional*, defaults to `-100`):
            Padding value to use for labels.
        max_prompt_length (`int` or `None`, *optional*, defaults to `512`):
            Maximum length of the prompt.
        max_completion_length (`int` or `None`, *optional*, defaults to `None`):
            Maximum length of the completion.
        max_length (`int` or `None`, *optional*, defaults to `1024`):
            Maximum length of the full sequence (prompt + completion).
        truncation_mode (`str`, *optional*, defaults to `"keep_end"`):
            Truncation mode to use when the sequence exceeds `max_length`. Possible values are `"keep_end"` and
            `"keep_start"`.
        padding_free (`bool`, *optional*, defaults to `False`):
            Whether forward passes are performed without padding by flattening all sequences in the batch
            into a single continuous sequence. This approach requires associating a `position_ids` vector to track
            positional information. Currently, this is only supported with the `flash_attention_2` mechanism, as it
            can handle the flattened batch structure.
        precompute_ref_log_probs (`bool`, *optional*, defaults to `False`):
            Whether to precompute the log probabilities from the reference model. Setting this to `True` allows
            training without needing the reference model during training, which can help reduce GPU memory usage. If
            set to `False` (default), the reference model will be used during training to compute log probabilities
            on-the-fly.
        precompute_ref_batch_size (`int` or `None`, *optional*, defaults to `None`):
            Batch size to use when precomputing reference model log probabilities. This can be set higher than the
            training batch size to speed up preprocessing. If `None`, defaults to `per_device_train_batch_size` for
            training and `per_device_eval_batch_size` for evaluation.
        tools (`Optional[list[Union[dict, Callable]]]`, *optional*, defaults to `None`):
            List of tools (callable functions) that will be accessible to the model.
            If the template does not support function calling, this argument will have no effect.

        > Parameters that control the training

        learning_rate (`float`, *optional*, defaults to `1e-6`):
            Initial learning rate for [`AdamW`] optimizer. The default value replaces that of
            [`~transformers.TrainingArguments`].
        loss_type (`str`, *optional*, defaults to `"sigmoid"`):
            Type of loss to use. Possible values are:

                - `"sigmoid"`: sigmoid loss from the original [DPO](https://huggingface.co/papers/2305.18290) paper.
                - `"hinge"`: hinge loss on the normalized likelihood from the [SLiC](https://huggingface.co/papers/2305.10425) paper.
                - `"ipo"`: IPO loss from the [IPO](https://huggingface.co/papers/2310.12036) paper.
                - `"exo_pair"`: pairwise EXO loss from the [EXO](https://huggingface.co/papers/2402.00856) paper.
                - `"nca_pair"`: pairwise NCA loss from the [NCA](https://huggingface.co/papers/2402.05369) paper.
                - `"robust"`: unbiased estimate of the DPO loss that is robust to preference noise from the [Robust DPO](https://huggingface.co/papers/2403.00409) paper.
                - `"bco_pair"`: pairwise BCO loss from the [BCO](https://huggingface.co/papers/2404.04656) paper.
                - `"sppo_hard"`: SPPO loss with hard label from the [SPPO](https://huggingface.co/papers/2405.00675) paper.
                - `"aot"`: AOT loss for paired datasets from the [AOT](https://huggingface.co/papers/2406.05882) paper.
                - `"aot_pair"`: AOT loss for unpaired datasets from the [AOT](https://huggingface.co/papers/2406.05882) paper.
                - `"discopop"`: DiscoPOP (a.k.a Log-Ratio Modulated Loss, LRML) loss from the [DiscoPOP](https://huggingface.co/papers/2406.08414) paper.
                - `"apo_zero"`: APO-zero loss from the [APO](https://huggingface.co/papers/2408.06266) paper.
                - `"apo_down"`: APO-down loss from the [APO](https://huggingface.co/papers/2408.06266) paper.

        beta (`float`, *optional*, defaults to `0.1`):
            Parameter controlling the deviation from the reference model. Higher β means less deviation from the
            reference model. For the IPO loss (`loss_type="ipo"`), β is the regularization parameter denoted by τ in
            the [paper](https://huggingface.co/papers/2310.12036).
        f_divergence_type (`str`, *optional*, defaults to `FDivergenceType.REVERSE_KL`):
            Type of f-divergence regularization function to compute divergence between policy and reference model.
        f_alpha_divergence_coef (`float`, *optional*, defaults to `1.0`):
            α coefficient in the α-divergence u^-α regularization function for DPO loss.
        reference_free (`bool`, *optional*, defaults to `False`):
            Whether to ignore the provided reference model and implicitly use a reference model that assigns equal
            probability to all responses.
        label_smoothing (`float`, *optional*, defaults to `0.0`):
            Robust DPO label smoothing parameter from the [cDPO](https://ericmitchell.ai/cdpo.pdf) report and
            [Robust DPO](https://huggingface.co/papers/2403.00409) paper that should be between `0.0` and `0.5`.
        use_weighting (`bool`, *optional*, defaults to `False`):
            Whether to weight the loss as done in the [WPO](https://huggingface.co/papers/2406.11827) paper.
        rpo_alpha (`float`, *optional*, defaults to `None`):
            α parameter from the [RPO](https://huggingface.co/papers/2404.19733) paper (v3), which controls the
            weighting of the NLL term in the loss. If `None`, no weighting is applied and the loss is the same as the
            DPO loss. The paper recommends `rpo_alpha=1.0`.
        discopop_tau (`float`, *optional*, defaults to `0.05`):
            τ/temperature parameter from the [DiscoPOP](https://huggingface.co/papers/2406.08414) paper, which controls
            the shape of log ratio modulated loss. The paper recommends the default value `discopop_tau=0.05`.
        sync_ref_model (`bool`, *optional*, defaults to `False`):
            Whether to synchronize the reference model with the active model every `ref_model_sync_steps` steps, using
            the `ref_model_mixup_alpha` parameter. This synchronization originites from the
            [TR-DPO](https://huggingface.co/papers/2404.09656) paper.
        ref_model_mixup_alpha (`float`, *optional*, defaults to `0.9`):
            α parameter from the [TR-DPO](https://huggingface.co/papers/2404.09656) paper, which controls the mix
            between the current policy and the previous reference policy during updates. The reference policy is
            updated according to the equation: `π_ref = α * π_θ + (1 - α) * π_ref_prev`. To use this parameter, you
            must set `sync_ref_model=True`.
        ref_model_sync_steps (`int`, *optional*, defaults to `64`):
            τ parameter from the [TR-DPO](https://huggingface.co/papers/2404.09656) paper, which determines how
            frequently the current policy is synchronized with the reference policy. To use this parameter, you must
            set `sync_ref_model=True`.

        > Parameters that control the logging

        generate_during_eval (`bool`, *optional*, defaults to `False`):
            Whether to generate and log completions from both the model and the reference model to W&B or Comet during
            evaluation.
    
    """
    vllm_sampling_params: Optional[Any] = field(
        default = None,
        metadata = {'help': 'vLLM SamplingParams'},
    )
    unsloth_num_chunks : Optional[int] = field(
        default = -1,
        metadata = {'help': 'Chunk size to reduce memory usage. -1 is most efficient.'},
    )
    def __init__(
        self,
        output_dir = None,
        overwrite_output_dir = None,
        do_train = False,
        do_eval = False,
        do_predict = False,
        eval_strategy = 'no',
        prediction_loss_only = False,
        per_device_train_batch_size = 4,
        per_device_eval_batch_size = 4,
        per_gpu_train_batch_size = None,
        per_gpu_eval_batch_size = None,
        gradient_accumulation_steps = 2,
        eval_accumulation_steps = 2,
        eval_delay = 0,
        torch_empty_cache_steps = 250,
        learning_rate = 5e-05,
        weight_decay = 0.01,
        adam_beta1 = 0.9,
        adam_beta2 = 0.999,
        adam_epsilon = 1e-08,
        max_grad_norm = 1.0,
        num_train_epochs = 3.0,
        max_steps = -1,
        lr_scheduler_type = 'linear',
        warmup_ratio = 0.1,
        warmup_steps = 0,
        log_level = 'passive',
        log_level_replica = 'warning',
        log_on_each_node = True,
        logging_dir = None,
        logging_strategy = 'steps',
        logging_first_step = False,
        logging_steps = 1,
        logging_nan_inf_filter = False,
        save_strategy = 'steps',
        save_steps = 500,
        save_total_limit = None,
        save_safetensors = True,
        save_on_each_node = False,
        save_only_model = False,
        restore_callback_states_from_checkpoint = False,
        no_cuda = False,
        use_cpu = False,
        use_mps_device = False,
        seed = 3407,
        data_seed = 3407,
        jit_mode_eval = False,
        use_ipex = False,
        bf16 = False,
        fp16 = False,
        fp16_opt_level = 'O1',
        half_precision_backend = 'auto',
        bf16_full_eval = False,
        fp16_full_eval = False,
        tf32 = None,
        local_rank = -1,
        ddp_backend = None,
        tpu_num_cores = None,
        tpu_metrics_debug = False,
        debug = '',
        dataloader_drop_last = False,
        eval_steps = None,
        dataloader_num_workers = 0,
        dataloader_prefetch_factor = None,
        past_index = -1,
        run_name = None,
        disable_tqdm = None,
        remove_unused_columns = True,
        label_names = None,
        load_best_model_at_end = False,
        metric_for_best_model = None,
        greater_is_better = None,
        ignore_data_skip = False,
        fsdp = '',
        fsdp_min_num_params = 0,
        fsdp_config = None,
        fsdp_transformer_layer_cls_to_wrap = None,
        accelerator_config = None,
        deepspeed = None,
        label_smoothing_factor = 0.0,
        optim = 'adamw_8bit',
        optim_args = None,
        adafactor = False,
        group_by_length = False,
        length_column_name = 'length',
        report_to = None,
        ddp_find_unused_parameters = None,
        ddp_bucket_cap_mb = None,
        ddp_broadcast_buffers = None,
        dataloader_pin_memory = True,
        dataloader_persistent_workers = False,
        skip_memory_metrics = True,
        use_legacy_prediction_loop = False,
        push_to_hub = False,
        resume_from_checkpoint = None,
        hub_model_id = None,
        hub_strategy = 'every_save',
        hub_token = None,
        hub_private_repo = None,
        hub_always_push = False,
        gradient_checkpointing = False,
        gradient_checkpointing_kwargs = None,
        include_inputs_for_metrics = False,
        eval_do_concat_batches = True,
        fp16_backend = 'auto',
        evaluation_strategy = None,
        push_to_hub_model_id = None,
        push_to_hub_organization = None,
        push_to_hub_token = None,
        mp_parameters = '',
        auto_find_batch_size = False,
        full_determinism = False,
        torchdynamo = None,
        ray_scope = 'last',
        ddp_timeout = 1800,
        torch_compile = False,
        torch_compile_backend = None,
        torch_compile_mode = None,
        dispatch_batches = None,
        split_batches = None,
        include_tokens_per_second = False,
        include_num_input_tokens_seen = False,
        neftune_noise_alpha = None,
        optim_target_modules = None,
        batch_eval_metrics = False,
        eval_on_start = False,
        use_liger_kernel = False,
        eval_use_gather_object = False,
        average_tokens_across_devices = False,
        model_init_kwargs = None,
        ref_model_init_kwargs = None,
        model_adapter_name = None,
        ref_adapter_name = None,
        force_use_ref_model = False,
        disable_dropout = True,
        use_logits_to_keep = False,
        dataset_num_proc = None,
        padding_value = None,
        label_pad_token_id = -100,
        max_prompt_length = 512,
        max_completion_length = None,
        max_length = 1024,
        truncation_mode = 'keep_end',
        padding_free = False,
        precompute_ref_log_probs = False,
        precompute_ref_batch_size = None,
        tools = None,
        loss_type = 'sigmoid',
        beta = 0.1,
        f_alpha_divergence_coef = 1.0,
        reference_free = False,
        label_smoothing = 0.0,
        use_weighting = False,
        rpo_alpha = None,
        discopop_tau = 0.05,
        sync_ref_model = False,
        ref_model_mixup_alpha = 0.9,
        ref_model_sync_steps = 64,
        generate_during_eval = False,
        use_num_logits_to_keep = False,
        vllm_sampling_params = None,
        unsloth_num_chunks = -1,
        **kwargs,
    ):
        if learning_rate < 1e-7: raise FloatingPointError(f'Unsloth: Your learning rate of `{learning_rate}` is too small and less than 1e-7! Consider increasing it, otherwise gradient updates will be close to 0!')
        if learning_rate > 1: raise OverflowError(f'Unsloth: Your learning rate of `{learning_rate}` is way too larger > 1! Consider decreasing it to 1e-1, otherwise gradient updates will explode!')
        if output_dir is None and save_strategy == 'steps' and save_steps == 500:
            output_dir = 'unsloth_training_checkpoints'
            save_strategy = 'no'
        if dataset_num_proc is None:
            from multiprocessing import cpu_count
            dataset_num_proc = cpu_count()
        
        super().__init__(
            output_dir = output_dir,
            overwrite_output_dir = overwrite_output_dir,
            do_train = do_train,
            do_eval = do_eval,
            do_predict = do_predict,
            eval_strategy = eval_strategy,
            prediction_loss_only = prediction_loss_only,
            per_device_train_batch_size = per_device_train_batch_size,
            per_device_eval_batch_size = per_device_eval_batch_size,
            per_gpu_train_batch_size = per_gpu_train_batch_size,
            per_gpu_eval_batch_size = per_gpu_eval_batch_size,
            gradient_accumulation_steps = gradient_accumulation_steps,
            eval_accumulation_steps = eval_accumulation_steps,
            eval_delay = eval_delay,
            torch_empty_cache_steps = torch_empty_cache_steps,
            learning_rate = learning_rate,
            weight_decay = weight_decay,
            adam_beta1 = adam_beta1,
            adam_beta2 = adam_beta2,
            adam_epsilon = adam_epsilon,
            max_grad_norm = max_grad_norm,
            num_train_epochs = num_train_epochs,
            max_steps = max_steps,
            lr_scheduler_type = lr_scheduler_type,
            warmup_ratio = warmup_ratio,
            warmup_steps = warmup_steps,
            log_level = log_level,
            log_level_replica = log_level_replica,
            log_on_each_node = log_on_each_node,
            logging_dir = logging_dir,
            logging_strategy = logging_strategy,
            logging_first_step = logging_first_step,
            logging_steps = logging_steps,
            logging_nan_inf_filter = logging_nan_inf_filter,
            save_strategy = save_strategy,
            save_steps = save_steps,
            save_total_limit = save_total_limit,
            save_safetensors = save_safetensors,
            save_on_each_node = save_on_each_node,
            save_only_model = save_only_model,
            restore_callback_states_from_checkpoint = restore_callback_states_from_checkpoint,
            no_cuda = no_cuda,
            use_cpu = use_cpu,
            use_mps_device = use_mps_device,
            seed = seed,
            data_seed = data_seed,
            jit_mode_eval = jit_mode_eval,
            use_ipex = use_ipex,
            bf16 = bf16,
            fp16 = fp16,
            fp16_opt_level = fp16_opt_level,
            half_precision_backend = half_precision_backend,
            bf16_full_eval = bf16_full_eval,
            fp16_full_eval = fp16_full_eval,
            tf32 = tf32,
            local_rank = local_rank,
            ddp_backend = ddp_backend,
            tpu_num_cores = tpu_num_cores,
            tpu_metrics_debug = tpu_metrics_debug,
            debug = debug,
            dataloader_drop_last = dataloader_drop_last,
            eval_steps = eval_steps,
            dataloader_num_workers = dataloader_num_workers,
            dataloader_prefetch_factor = dataloader_prefetch_factor,
            past_index = past_index,
            run_name = run_name,
            disable_tqdm = disable_tqdm,
            remove_unused_columns = remove_unused_columns,
            label_names = label_names,
            load_best_model_at_end = load_best_model_at_end,
            metric_for_best_model = metric_for_best_model,
            greater_is_better = greater_is_better,
            ignore_data_skip = ignore_data_skip,
            fsdp = fsdp,
            fsdp_min_num_params = fsdp_min_num_params,
            fsdp_config = fsdp_config,
            fsdp_transformer_layer_cls_to_wrap = fsdp_transformer_layer_cls_to_wrap,
            accelerator_config = accelerator_config,
            deepspeed = deepspeed,
            label_smoothing_factor = label_smoothing_factor,
            optim = optim,
            optim_args = optim_args,
            adafactor = adafactor,
            group_by_length = group_by_length,
            length_column_name = length_column_name,
            report_to = report_to,
            ddp_find_unused_parameters = ddp_find_unused_parameters,
            ddp_bucket_cap_mb = ddp_bucket_cap_mb,
            ddp_broadcast_buffers = ddp_broadcast_buffers,
            dataloader_pin_memory = dataloader_pin_memory,
            dataloader_persistent_workers = dataloader_persistent_workers,
            skip_memory_metrics = skip_memory_metrics,
            use_legacy_prediction_loop = use_legacy_prediction_loop,
            push_to_hub = push_to_hub,
            resume_from_checkpoint = resume_from_checkpoint,
            hub_model_id = hub_model_id,
            hub_strategy = hub_strategy,
            hub_token = hub_token,
            hub_private_repo = hub_private_repo,
            hub_always_push = hub_always_push,
            gradient_checkpointing = gradient_checkpointing,
            gradient_checkpointing_kwargs = gradient_checkpointing_kwargs,
            include_inputs_for_metrics = include_inputs_for_metrics,
            eval_do_concat_batches = eval_do_concat_batches,
            fp16_backend = fp16_backend,
            evaluation_strategy = evaluation_strategy,
            push_to_hub_model_id = push_to_hub_model_id,
            push_to_hub_organization = push_to_hub_organization,
            push_to_hub_token = push_to_hub_token,
            mp_parameters = mp_parameters,
            auto_find_batch_size = auto_find_batch_size,
            full_determinism = full_determinism,
            torchdynamo = torchdynamo,
            ray_scope = ray_scope,
            ddp_timeout = ddp_timeout,
            torch_compile = torch_compile,
            torch_compile_backend = torch_compile_backend,
            torch_compile_mode = torch_compile_mode,
            dispatch_batches = dispatch_batches,
            split_batches = split_batches,
            include_tokens_per_second = include_tokens_per_second,
            include_num_input_tokens_seen = include_num_input_tokens_seen,
            neftune_noise_alpha = neftune_noise_alpha,
            optim_target_modules = optim_target_modules,
            batch_eval_metrics = batch_eval_metrics,
            eval_on_start = eval_on_start,
            use_liger_kernel = use_liger_kernel,
            eval_use_gather_object = eval_use_gather_object,
            average_tokens_across_devices = average_tokens_across_devices,
            model_init_kwargs = model_init_kwargs,
            ref_model_init_kwargs = ref_model_init_kwargs,
            model_adapter_name = model_adapter_name,
            ref_adapter_name = ref_adapter_name,
            force_use_ref_model = force_use_ref_model,
            disable_dropout = disable_dropout,
            use_logits_to_keep = use_logits_to_keep,
            dataset_num_proc = dataset_num_proc,
            padding_value = padding_value,
            label_pad_token_id = label_pad_token_id,
            max_prompt_length = max_prompt_length,
            max_completion_length = max_completion_length,
            max_length = max_length,
            truncation_mode = truncation_mode,
            padding_free = padding_free,
            precompute_ref_log_probs = precompute_ref_log_probs,
            precompute_ref_batch_size = precompute_ref_batch_size,
            tools = tools,
            loss_type = loss_type,
            beta = beta,
            f_alpha_divergence_coef = f_alpha_divergence_coef,
            reference_free = reference_free,
            label_smoothing = label_smoothing,
            use_weighting = use_weighting,
            rpo_alpha = rpo_alpha,
            discopop_tau = discopop_tau,
            sync_ref_model = sync_ref_model,
            ref_model_mixup_alpha = ref_model_mixup_alpha,
            ref_model_sync_steps = ref_model_sync_steps,
            generate_during_eval = generate_during_eval,
            use_num_logits_to_keep = use_num_logits_to_keep,**kwargs)
        self.vllm_sampling_params = vllm_sampling_params
        self.unsloth_num_chunks = unsloth_num_chunks
pass

class _UnslothDPOTrainer(Trainer):
    r""""""

    _tag_names = ["trl", "dpo"]

    @deprecate_kwarg(
        "tokenizer", "0.16.0", "processing_class", warn_if_greater_or_equal_version=True, raise_if_both_names=True
    )
    def __init__(
        self,
        model: Optional[Union[PreTrainedModel, nn.Module, str]] = None,
        ref_model: Optional[Union[PreTrainedModel, nn.Module, str]] = None,
        args: Optional[DPOConfig] = None,
        data_collator: Optional[DataCollator] = None,
        train_dataset: Optional[Dataset] = None,
        eval_dataset: Optional[Union[Dataset, dict[str, Dataset]]] = None,
        processing_class: Optional[
            Union[PreTrainedTokenizerBase, BaseImageProcessor, FeatureExtractionMixin, ProcessorMixin]
        ] = None,
        model_init: Optional[Callable[[], PreTrainedModel]] = None,
        compute_metrics: Optional[Callable[[EvalLoopOutput], dict]] = None,
        callbacks: Optional[list[TrainerCallback]] = None,
        optimizers: tuple[torch.optim.Optimizer, torch.optim.lr_scheduler.LambdaLR] = (None, None),
        preprocess_logits_for_metrics: Optional[Callable[[torch.Tensor, torch.Tensor], torch.Tensor]] = None,
        peft_config: Optional[dict] = None,
    ):
        if model is None:
            raise ValueError("No model provided. Please provide a model to train.")

        if not isinstance(model, str) and ref_model is model:
            raise ValueError(
                "`model` and `ref_model` cannot be the same object. If you want `ref_model` to be the "
                "same as `model`, you must mass a copy of it, or `None` if you use peft."
            )

        if args.model_init_kwargs is None:
            model_init_kwargs = {}
        elif not isinstance(model, str):
            raise ValueError(
                "You passed model_init_kwargs to the DPOTrainer/DPOConfig, but your model is already instantiated."
            )
        else:
            model_init_kwargs = args.model_init_kwargs
            torch_dtype = model_init_kwargs.get("torch_dtype")
            if torch_dtype is not None:
                # Convert to `torch.dtype` if an str is passed
                if isinstance(torch_dtype, str) and torch_dtype != "auto":
                    torch_dtype = getattr(torch, torch_dtype)
                if torch_dtype != "auto" and not isinstance(torch_dtype, torch.dtype):
                    raise ValueError(
                        f"Invalid `torch_dtype` passed to the DPOConfig. Expected a string with either `torch.dtype` or 'auto', but got {torch_dtype}."
                    )
                model_init_kwargs["torch_dtype"] = torch_dtype

        if args.ref_model_init_kwargs is None:
            ref_model_init_kwargs = {}
        elif not isinstance(ref_model, str):
            raise ValueError(
                "You passed ref_model_init_kwargs to the DPOTrainer/DPOConfig, but your ref_model is already instantiated."
            )
        else:
            ref_model_init_kwargs = args.ref_model_init_kwargs
            torch_dtype = ref_model_init_kwargs.get("torch_dtype")
            if torch_dtype is not None:
                # Convert to `torch.dtype` if an str is passed
                if isinstance(torch_dtype, str) and torch_dtype != "auto":
                    torch_dtype = getattr(torch, torch_dtype)
                if torch_dtype != "auto" and not isinstance(torch_dtype, torch.dtype):
                    raise ValueError(
                        f"Invalid `torch_dtype` passed to the DPOConfig. Expected a string with either `torch.dtype` or 'auto', but got {torch_dtype}."
                    )
                ref_model_init_kwargs["torch_dtype"] = torch_dtype

        if isinstance(model, str):
            model = AutoModelForCausalLM.from_pretrained(model, **model_init_kwargs)

        if isinstance(ref_model, str):
            ref_model = AutoModelForCausalLM.from_pretrained(ref_model, **ref_model_init_kwargs)

        # Initialize this variable to False. This helps tracking the case when `peft_module_casting_to_bf16`
        # has been called in order to properly call autocast if needed.
        self._peft_has_been_casted_to_bf16 = False

        if not is_peft_available() and peft_config is not None:
            raise ValueError(
                "PEFT is not installed and you passed a `peft_config` in the trainer's kwargs, please install it to use the PEFT models"
            )
        elif is_peft_available() and peft_config is not None:
            # if model is a peft model and we have a peft_config, we merge and unload it first
            if isinstance(model, PeftModel):
                model = model.merge_and_unload()

            if ref_model is not None and not args.force_use_ref_model:
                raise ValueError(
                    "You passed both a ref_model and a peft_config. For training PEFT adapters with DPO there is no need to pass a reference"
                    " model. Please pass `ref_model=None` in case you want to train PEFT adapters, or pass a ref_model with `force_use_ref_model=True` in DPOTrainer's init."
                    " if you want to use a different ref_model."
                )

            if getattr(model, "is_loaded_in_8bit", False) or getattr(model, "is_loaded_in_4bit", False):
                _support_gc_kwargs = hasattr(
                    args, "gradient_checkpointing_kwargs"
                ) and "gradient_checkpointing_kwargs" in list(
                    inspect.signature(prepare_model_for_kbit_training).parameters
                )

                prepare_model_kwargs = {"use_gradient_checkpointing": args.gradient_checkpointing}

                if _support_gc_kwargs:
                    prepare_model_kwargs["gradient_checkpointing_kwargs"] = args.gradient_checkpointing_kwargs

                model = prepare_model_for_kbit_training(model, **prepare_model_kwargs)
            elif getattr(args, "gradient_checkpointing", False):
                # For backward compatibility with older versions of transformers
                if hasattr(model, "enable_input_require_grads"):
                    model.enable_input_require_grads()
                else:

                    def make_inputs_require_grad(module, input, output):
                        output.requires_grad_(True)

                    model.get_input_embeddings().register_forward_hook(make_inputs_require_grad)

            # get peft model with the given config
            model = model
            if args.bf16 and getattr(model, "is_loaded_in_4bit", False):
                peft_module_casting_to_bf16(model)
                # If args.bf16 we need to explicitly call `generate` with torch amp autocast context manager
                self._peft_has_been_casted_to_bf16 = True

        # For models that use gradient_checkpointing, we need to attach a hook that enables input
        # to explicitly have `requires_grad=True`, otherwise training will either silently
        # fail or completely fail.
        elif getattr(args, "gradient_checkpointing", False):
            # For backward compatibility with older versions of transformers
            if hasattr(model, "enable_input_require_grads"):
                model.enable_input_require_grads()
            else:

                def make_inputs_require_grad(module, input, output):
                    output.requires_grad_(True)

                model.get_input_embeddings().register_forward_hook(make_inputs_require_grad)

        if args.generate_during_eval and not (is_wandb_available() or is_comet_available()):
            raise ValueError(
                "`generate_during_eval=True` requires Weights and Biases or Comet to be installed."
                " Please install `wandb` or `comet-ml` to resolve."
            )

        self.is_encoder_decoder = model.config.is_encoder_decoder
        self.is_vision_model = model.config.model_type in MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES.keys()
        self.is_peft_model = is_peft_available() and isinstance(model, PeftModel)
        self.model_adapter_name = args.model_adapter_name
        self.ref_adapter_name = args.ref_adapter_name
        self.reference_free = args.reference_free

        if ref_model:
            self.ref_model = ref_model
        elif self.is_peft_model or args.precompute_ref_log_probs:
            # The `model` with adapters turned off will be used as the reference model
            self.ref_model = None
        else:
            self.ref_model = create_reference_model(model)

        if processing_class is None:
            raise ValueError("processing_class must be specified to tokenize a DPO dataset.")

        if args.padding_value is not None:
            self.padding_value = args.padding_value
        else:
            if hasattr(processing_class, "pad_token_id") and processing_class.pad_token_id is not None:
                self.padding_value = processing_class.pad_token_id
            elif hasattr(processing_class, "tokenizer") and processing_class.tokenizer.pad_token_id is not None:
                self.padding_value = processing_class.tokenizer.pad_token_id
            else:
                raise ValueError(
                    "`padding_value` is not specified in `DPOConfig`, and `pad_token_id` is missing in the "
                    "`processing_class`. Please either set the `padding_value` argument in `DPOConfig`, or set "
                    "`tokenizer.pad_token` (e.g., `tokenizer.pad_token = tokenizer.eos_token`) before instantiating "
                    "the trainer."
                )

        if data_collator is None:
            data_collator = DataCollatorForPreference(pad_token_id=self.padding_value)

        # Disable dropout in the model and reference model
        if args.disable_dropout:
            disable_dropout_in_model(model)
            if self.ref_model is not None:
                disable_dropout_in_model(self.ref_model)

        self.generate_during_eval = args.generate_during_eval
        self.label_pad_token_id = args.label_pad_token_id
        self.max_prompt_length = args.max_prompt_length
        self.max_completion_length = args.max_completion_length
        self.max_length = args.max_length
        self.truncation_mode = args.truncation_mode
        self.precompute_ref_log_probs = args.precompute_ref_log_probs
        self.use_logits_to_keep = args.use_logits_to_keep

        if args.padding_free:
            if model.config._attn_implementation != "flash_attention_2":
                warnings.warn(
                    "Padding-free training is enabled, but the attention implementation is not set to "
                    "'flash_attention_2'. Padding-free training flattens batches into a single sequence, and "
                    "'flash_attention_2' is the only known attention mechanism that reliably supports this. Using "
                    "other implementations may lead to unexpected behavior. To ensure compatibility, set "
                    "`attn_implementation='flash_attention_2'` in the model configuration, or verify that your "
                    "attention mechanism can handle flattened sequences."
                )
        self.padding_free = args.padding_free

        # Since ref_logs are precomputed on the first call to get_train/eval_dataloader
        # keep track of first called to avoid computation of future calls
        self._precomputed_train_ref_log_probs = False
        self._precomputed_eval_ref_log_probs = False

        if (
            args.loss_type in ["hinge", "ipo", "bco_pair", "sppo_hard", "nca_pair", "apo_zero", "apo_down"]
            and args.label_smoothing > 0
        ):
            warnings.warn(
                f"You are using the {args.loss_type} loss type that does not support label smoothing. The "
                "`label_smoothing` parameter will be ignored. Set `label_smoothing` to `0.0` to remove this warning.",
                UserWarning,
            )
        if args.loss_type == "kto_pair":
            raise ValueError("Support for kto_pair has been removed in DPOTrainer. Please use KTOTrainer.")

        self.beta = args.beta
        self.label_smoothing = args.label_smoothing
        self.loss_type = args.loss_type
        self.aux_loss_enabled = getattr(model.config, "output_router_logits", False)
        self.use_weighting = args.use_weighting
        self.aux_loss_coef = getattr(model.config, "router_aux_loss_coef", 0.0)
        if self.aux_loss_enabled and self.aux_loss_coef == 0.0:
            warnings.warn(
                "You set `output_router_logits` to `True` in the model config, but `router_aux_loss_coef` is set to "
                "`0.0`, meaning the auxiliary loss will not be used. Either set `router_aux_loss_coef` to a value "
                "greater than `0.0`, or set `output_router_logits` to `False` if you don't want to use the auxiliary "
                "loss.",
                UserWarning,
            )

        self._stored_metrics = defaultdict(lambda: defaultdict(list))
        self.f_divergence_type = args.f_divergence_type
        self.f_divergence_params = {FDivergenceConstants.ALPHA_DIVERGENCE_COEF_KEY: args.f_alpha_divergence_coef}
        self.dataset_num_proc = args.dataset_num_proc

        # The trainer estimates the number of FLOPs (floating-point operations) using the number of elements in the
        # input tensor associated with the key "input_ids". However, in DPO, the sampled data does not include the
        # "input_ids" key. Instead, the available keys are "prompt_input_ids", "chosen_input_ids", and
        # "rejected_input_ids". As a result, the trainer issues the warning: "Could not estimate the number of tokens
        # of the input, floating-point operations will not be computed." To suppress this warning, we set the
        # "estimate_tokens" key in the model's "warnings_issued" dictionary to True. This acts as a flag to indicate
        # that the warning has already been issued.
        model.warnings_issued["estimate_tokens"] = True

        # Dataset preparation
        train_dataset = self._prepare_dataset(train_dataset, processing_class, args, "train")
        if eval_dataset is not None:
            if isinstance(eval_dataset, dict):
                eval_dataset = {
                    key: self._prepare_dataset(dataset, processing_class, args, key)
                    for key, dataset in eval_dataset.items()
                }
            else:
                eval_dataset = self._prepare_dataset(eval_dataset, processing_class, args, "eval")

        super().__init__(
            model=model,
            args=args,
            data_collator=data_collator,
            train_dataset=train_dataset,
            eval_dataset=eval_dataset,
            processing_class=processing_class,
            model_init=model_init,
            compute_metrics=compute_metrics,
            callbacks=callbacks,
            optimizers=optimizers,
            preprocess_logits_for_metrics=preprocess_logits_for_metrics,
        )

        # Gradient accumulation requires scaled loss. Normally, loss scaling in the parent class depends on whether the
        # model accepts loss-related kwargs. Since we compute our own loss, this check is irrelevant. We set
        # self.model_accepts_loss_kwargs to False to enable scaling.
        self.model_accepts_loss_kwargs = False

        # Add tags for models that have been loaded with the correct transformers version
        if hasattr(self.model, "add_model_tags"):
            self.model.add_model_tags(self._tag_names)

        if not hasattr(self, "accelerator"):
            raise AttributeError(
                "Your `Trainer` does not have an `accelerator` object. Consider upgrading `transformers`."
            )

        # Deepspeed Zero-3 does not support precompute_ref_log_probs
        if self.is_deepspeed_enabled:
            if self.accelerator.state.deepspeed_plugin.zero_stage == 3 and self.precompute_ref_log_probs:
                raise ValueError(
                    "You cannot use `precompute_ref_log_probs=True` with Deepspeed ZeRO-3. Please set `precompute_ref_log_probs=False`."
                )

        if self.ref_model is None:
            if not (self.is_peft_model or self.precompute_ref_log_probs):
                raise ValueError(
                    "No reference model and model is not a Peft model. Try setting `precompute_ref_log_probs=True`"
                )
            if args.sync_ref_model:
                raise ValueError(
                    "You currently cannot use `ref_model=None` with TR-DPO method. Please provide `ref_model`."
                )
        else:
            if self.is_deepspeed_enabled:
                self.ref_model = self._prepare_deepspeed(self.ref_model)
            else:
                self.ref_model = self.accelerator.prepare_model(self.ref_model, evaluation_mode=True)

        if args.sync_ref_model:
            if self.precompute_ref_log_probs:
                raise ValueError(
                    "You cannot use `precompute_ref_log_probs=True` with TR-DPO method. Please set `precompute_ref_log_probs=False`."
                )

            self.add_callback(SyncRefModelCallback(ref_model=self.ref_model, accelerator=self.accelerator))

        if self.loss_type == "bco_pair":
            self.running = RunningMoments(self.accelerator)

    def _prepare_dataset(
        self,
        dataset: Union[Dataset, IterableDataset],
        processing_class: Union[PreTrainedTokenizerBase, BaseImageProcessor, FeatureExtractionMixin, ProcessorMixin],
        args: DPOConfig,
        dataset_name: str,
    ) -> Union[Dataset, IterableDataset]:
        # Build the kwargs for the `map` function
        map_kwargs = {"writer_batch_size": 10}
        if isinstance(dataset, Dataset):  # IterableDataset does not support num_proc
            map_kwargs["num_proc"] = args.dataset_num_proc

        with PartialState().local_main_process_first():
            # Extract prompt if needed
            if isinstance(dataset, Dataset):  # `IterableDataset.map` does not support `desc`
                map_kwargs["desc"] = f"Extracting prompt in {dataset_name} dataset"
            dataset = dataset.map(maybe_extract_prompt, **map_kwargs)

            # Apply the chat template if needed
            if isinstance(dataset, Dataset):  # `IterableDataset.map` does not support `desc`
                map_kwargs["desc"] = f"Applying chat template to {dataset_name} dataset"
            dataset = dataset.map(
                maybe_apply_chat_template, fn_kwargs={"tokenizer": processing_class, "tools": args.tools}, **map_kwargs
            )

            # Tokenize the dataset
            if isinstance(dataset, Dataset):  # `IterableDataset.map` does not support `desc`
                map_kwargs["desc"] = f"Tokenizing {dataset_name} dataset"

            dataset = dataset.map(
                self.tokenize_row if not self.is_vision_model else self.process_row,
                remove_columns=["prompt", "chosen", "rejected"],
                fn_kwargs={
                    "processing_class": processing_class,
                    "max_prompt_length": args.max_prompt_length,
                    "max_completion_length": args.max_completion_length,
                    # for enc-dec, we add the special tokens ([bos_token] + prompt + [eos_token]; completion + [eos_token])
                    "add_special_tokens": False,
                },
                **map_kwargs,
            )

        return dataset

    @staticmethod
    def tokenize_row(features, processing_class, max_prompt_length, max_completion_length, add_special_tokens):
        """
        Tokenize a row of the dataset.

        Args:
            features (`dict[str, str]`):
                Row of the dataset, should contain the keys `"prompt"`, `"chosen"`, and `"rejected"`.
            processing_class (`PreTrainedTokenizerBase`):
                Processing class used to process the data.
            max_prompt_length (`int` or `None`):
                Maximum length of the prompt sequence. If `None`, the prompt sequence is not truncated.
            max_completion_length (`int` or `None`):
                Maximum length of the completion sequences. If `None`, the completion sequences are not truncated.
            add_special_tokens (`bool`):
                Whether to add special tokens to the sequences. Typically used for encoder-decoder models. If `True`,
                the prompt sequence will have a bos token prepended and an eos token appended. In any case, the
                completion sequences will have an eos token appended.

        Returns:
            `dict[str, list[int]]`:
                Tokenized sequences with the keys `"prompt_input_ids"`, `"chosen_input_ids"`, and
                `"rejected_input_ids".

        Example:
        ```python
        >>> from transformers import GPT2Tokenizer
        >>> tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
        >>> features = {"prompt": "The sky is", "chosen": " blue", "rejected": " green"}
        >>> DPOTrainer.tokenize_row(
        ...     features, tokenizer, max_prompt_length=3, max_completion_length=3, add_special_tokens=False
        ... )
        {'prompt_input_ids': [464, 6766, 318], 'chosen_input_ids': [4171, 50256], 'rejected_input_ids': [4077, 50256]}
        ```
        """
        tokenizer = processing_class  # the processing class is a tokenizer
        prompt_input_ids = tokenizer(features["prompt"], add_special_tokens=False)["input_ids"]
        chosen_input_ids = tokenizer(features["chosen"], add_special_tokens=False)["input_ids"]
        rejected_input_ids = tokenizer(features["rejected"], add_special_tokens=False)["input_ids"]

        # Add special tokens (typically for encoder-decoder models)
        if add_special_tokens:
            if tokenizer.bos_token_id is not None:
                prompt_input_ids = [tokenizer.bos_token_id] + prompt_input_ids
            if tokenizer.eos_token_id is not None:
                prompt_input_ids = prompt_input_ids + [tokenizer.eos_token_id]
        chosen_input_ids = chosen_input_ids + [tokenizer.eos_token_id]
        rejected_input_ids = rejected_input_ids + [tokenizer.eos_token_id]

        # Truncate prompt and completion sequences
        if max_prompt_length is not None:
            prompt_input_ids = prompt_input_ids[-max_prompt_length:]
        if max_completion_length is not None:
            chosen_input_ids = chosen_input_ids[:max_completion_length]
            rejected_input_ids = rejected_input_ids[:max_completion_length]

        return {
            "prompt_input_ids": prompt_input_ids,
            "chosen_input_ids": chosen_input_ids,
            "rejected_input_ids": rejected_input_ids,
        }

    @staticmethod
    def process_row(features, processing_class, max_prompt_length, max_completion_length, add_special_tokens):
        """
        Same as `tokenize_row` but for vision models. Please refer to `tokenize_row` for more information.
        """
        processor, tokenizer = processing_class, processing_class.tokenizer  # the processing class is a processor
        processed_features = processor(images=features["images"], text=features["prompt"], add_special_tokens=False)

        prompt_input_ids = processed_features["input_ids"][0]
        pixel_values = processed_features["pixel_values"][0]
        chosen_input_ids = tokenizer(features["chosen"], add_special_tokens=False)["input_ids"]
        rejected_input_ids = tokenizer(features["rejected"], add_special_tokens=False)["input_ids"]

        # Add special tokens (typically for encoder-decoder models)
        if add_special_tokens:
            if tokenizer.bos_token_id is not None:
                prompt_input_ids = [tokenizer.bos_token_id] + prompt_input_ids
            if tokenizer.eos_token_id is not None:
                prompt_input_ids = prompt_input_ids + [tokenizer.eos_token_id]
        chosen_input_ids = chosen_input_ids + [tokenizer.eos_token_id]
        rejected_input_ids = rejected_input_ids + [tokenizer.eos_token_id]

        # Truncate prompt and completion sequences
        if max_prompt_length is not None:
            prompt_input_ids = prompt_input_ids[-max_prompt_length:]
        if max_completion_length is not None:
            chosen_input_ids = chosen_input_ids[:max_completion_length]
            rejected_input_ids = rejected_input_ids[:max_completion_length]

        output = {
            "prompt_input_ids": prompt_input_ids,
            "pixel_values": pixel_values,
            "chosen_input_ids": chosen_input_ids,
            "rejected_input_ids": rejected_input_ids,
        }

        if "pixel_attention_mask" in processed_features:
            output["pixel_attention_mask"] = processed_features["pixel_attention_mask"][0]
        if "image_sizes" in processed_features:
            output["image_sizes"] = processed_features["image_sizes"][0]

        return output

    def _prepare_deepspeed(self, model: PreTrainedModelWrapper):
        # Adapted from accelerate: https://github.com/huggingface/accelerate/blob/739b135f8367becb67ffaada12fe76e3aa60fefd/src/accelerate/accelerator.py#L1473
        deepspeed_plugin = self.accelerator.state.deepspeed_plugin
        config_kwargs = deepcopy(deepspeed_plugin.deepspeed_config)

        if model is not None:
            if hasattr(model, "config"):
                hidden_size = (
                    max(model.config.hidden_sizes)
                    if getattr(model.config, "hidden_sizes", None)
                    else getattr(model.config, "hidden_size", None)
                )
                if hidden_size is not None and config_kwargs["zero_optimization"]["stage"] == 3:
                    # Note that `stage3_prefetch_bucket_size` can produce DeepSpeed messages like: `Invalidate trace cache @ step 0: expected module 1, but got module 0`
                    # This is expected and is not an error, see: https://github.com/microsoft/DeepSpeed/discussions/4081
                    config_kwargs.update(
                        {
                            "zero_optimization.reduce_bucket_size": hidden_size * hidden_size,
                            "zero_optimization.stage3_param_persistence_threshold": 10 * hidden_size,
                            "zero_optimization.stage3_prefetch_bucket_size": 0.9 * hidden_size * hidden_size,
                        }
                    )

        # If ZeRO-3 is used, we shard both the active and reference model.
        # Otherwise, we assume the reference model fits in memory and is initialized on each device with ZeRO disabled (stage 0)
        if config_kwargs["zero_optimization"]["stage"] != 3:
            config_kwargs["zero_optimization"]["stage"] = 0
        model, *_ = deepspeed.initialize(model=model, config=config_kwargs)
        model.eval()
        return model

    def _set_signature_columns_if_needed(self):
        # If `self.args.remove_unused_columns` is True, non-signature columns are removed.
        # By default, this method sets `self._signature_columns` to the model's expected inputs.
        # In DPOTrainer, we preprocess data, so using the model's signature columns doesn't work.
        # Instead, we set them to the columns expected by `DataCollatorForPreference`, hence the override.
        if self._signature_columns is None:
            self._signature_columns = [
                "prompt_input_ids",
                "chosen_input_ids",
                "rejected_input_ids",
                "image_sizes",
                "ref_chosen_logps",
                "ref_rejected_logps",
            ]

    def get_train_dataloader(self) -> DataLoader:
        """
        Returns the training [`~torch.utils.data.DataLoader`].

        Subclass of transformers.src.transformers.trainer.get_train_dataloader to precompute `ref_log_probs`.
        """

        if self.precompute_ref_log_probs and not self._precomputed_train_ref_log_probs:
            batch_size = self.args.precompute_ref_batch_size or self.args.per_device_train_batch_size
            dataloader_params = {
                "batch_size": batch_size,
                "collate_fn": self.data_collator,
                "num_workers": self.args.dataloader_num_workers,
                "pin_memory": self.args.dataloader_pin_memory,
                "shuffle": False,
            }

            # prepare dataloader
            data_loader = self.accelerator.prepare(DataLoader(self.train_dataset, **dataloader_params))

            ref_chosen_logps = []
            ref_rejected_logps = []
            for padded_batch in tqdm(iterable=data_loader, desc="Train dataset reference log probs"):
                ref_chosen_logp, ref_rejected_logp = self.compute_ref_log_probs(padded_batch)
                ref_chosen_logp, ref_rejected_logp = self.accelerator.gather_for_metrics(
                    (ref_chosen_logp, ref_rejected_logp)
                )
                ref_chosen_logps.append(ref_chosen_logp.cpu())
                ref_rejected_logps.append(ref_rejected_logp.cpu())

                # Unnecessary cache clearing to avoid OOM
                empty_cache()
                self.accelerator.free_memory()

            all_ref_chosen_logps = torch.cat(ref_chosen_logps).float().numpy()
            all_ref_rejected_logps = torch.cat(ref_rejected_logps).float().numpy()

            self.train_dataset = self.train_dataset.add_column(name="ref_chosen_logps", column=all_ref_chosen_logps)
            self.train_dataset = self.train_dataset.add_column(
                name="ref_rejected_logps", column=all_ref_rejected_logps
            )

            self._precomputed_train_ref_log_probs = True

        return super().get_train_dataloader()

    def get_eval_dataloader(self, eval_dataset: Optional[Dataset] = None) -> DataLoader:
        """
        Returns the evaluation [`~torch.utils.data.DataLoader`].

        Subclass of transformers.src.transformers.trainer.get_eval_dataloader to precompute `ref_log_probs`.

        Args:
            eval_dataset (`torch.utils.data.Dataset`, *optional*):
                If provided, will override `self.eval_dataset`. If it is a [`~datasets.Dataset`], columns not accepted
                by the `model.forward()` method are automatically removed. It must implement `__len__`.
        """
        if eval_dataset is None and self.eval_dataset is None:
            raise ValueError("Trainer: evaluation requires an eval_dataset.")
        eval_dataset = eval_dataset if eval_dataset is not None else self.eval_dataset

        if self.precompute_ref_log_probs and not self._precomputed_eval_ref_log_probs:
            batch_size = self.args.precompute_ref_batch_size or self.args.per_device_eval_batch_size
            dataloader_params = {
                "batch_size": batch_size,
                "collate_fn": self.data_collator,
                "num_workers": self.args.dataloader_num_workers,
                "pin_memory": self.args.dataloader_pin_memory,
                "shuffle": False,
            }

            # prepare dataloader
            data_loader = self.accelerator.prepare(DataLoader(eval_dataset, **dataloader_params))

            ref_chosen_logps = []
            ref_rejected_logps = []
            for padded_batch in tqdm(iterable=data_loader, desc="Eval dataset reference log probs"):
                ref_chosen_logp, ref_rejected_logp = self.compute_ref_log_probs(padded_batch)
                ref_chosen_logp, ref_rejected_logp = self.accelerator.gather_for_metrics(
                    (ref_chosen_logp, ref_rejected_logp)
                )
                ref_chosen_logps.append(ref_chosen_logp.cpu())
                ref_rejected_logps.append(ref_rejected_logp.cpu())

            all_ref_chosen_logps = torch.cat(ref_chosen_logps).float().numpy()
            all_ref_rejected_logps = torch.cat(ref_rejected_logps).float().numpy()

            eval_dataset = eval_dataset.add_column(name="ref_chosen_logps", column=all_ref_chosen_logps)
            eval_dataset = eval_dataset.add_column(name="ref_rejected_logps", column=all_ref_rejected_logps)

            # Save calculated ref_chosen_logps and ref_rejected_logps to the eval_dataset for subsequent runs
            if self.eval_dataset is not None:
                self.eval_dataset = eval_dataset
            self._precomputed_eval_ref_log_probs = True

        return super().get_eval_dataloader(eval_dataset=eval_dataset)

    @contextmanager
    def null_ref_context(self):
        """Context manager for handling null reference model (that is, peft adapter manipulation)."""
        with (
            self.accelerator.unwrap_model(self.model).disable_adapter()
            if self.is_peft_model and not self.ref_adapter_name
            else nullcontext()
        ):
            if self.ref_adapter_name:
                self.model.set_adapter(self.ref_adapter_name)
            yield
            if self.ref_adapter_name:
                self.model.set_adapter(self.model_adapter_name or "default")

    def compute_ref_log_probs(self, batch: dict[str, torch.LongTensor]) -> dict:
        """Computes log probabilities of the reference model for a single padded batch of a DPO specific dataset."""
        device_type = "xpu" if is_torch_xpu_available() else "cuda"
        compte_ref_context_manager = amp.autocast(device_type) if self._peft_has_been_casted_to_bf16 else nullcontext()
        with torch.no_grad(), compte_ref_context_manager:
            if self.ref_model is None:
                with self.null_ref_context():
                    ref_model_output = self.concatenated_forward(self.model, batch)
            else:
                ref_model_output = self.concatenated_forward(self.ref_model, batch)
        return ref_model_output["chosen_logps"], ref_model_output["rejected_logps"]

    @staticmethod
    def concatenated_inputs(
        batch: dict[str, Union[list, torch.LongTensor]], padding_value: int
    ) -> dict[str, torch.LongTensor]:
        """
        Concatenate the `chosen` and `rejected` inputs from the batch into a single tensor for both the prompt
        and completion sequences.

        Args:
            batch (`dict[str, Union[list, torch.LongTensor]]`):
                A batch of input data. The batch must contain the following keys:

                - `"prompt_input_ids"`: Tensor of shape `(batch_size, prompt_length)` representing the prompt input IDs.
                - `"chosen_input_ids"`: Tensor of shape `(batch_size, chosen_length)` representing the chosen completion input IDs.
                - `"rejected_input_ids"`: Tensor of shape `(batch_size, rejected_length)` representing the rejected completion input IDs.
                - `"prompt_pixel_values"` (optional): Tensor for pixel values, if available.
                - `"prompt_pixel_attention_mask"` (optional): Tensor for pixel attention masks, if available.

            padding_value (`int`):
                The padding value to use for the concatenated completion sequences (`chosen_input_ids` and
                `rejected_input_ids`).

        Returns:
            `dict[str, torch.LongTensor]`: A dictionary containing:

                - `"prompt_input_ids"`: Concatenated prompt input IDs of shape `(2 * batch_size, prompt_length)`.
                - `"completion_input_ids"`: Concatenated chosen and rejected completion input IDs of shape `(2 * batch_size, max_completion_length)`.
                - `"prompt_attention_mask"`: Concatenated prompt attention masks of shape `(2 * batch_size, prompt_length)`.
                - `"completion_attention_mask"`: Concatenated chosen and rejected attention masks of shape `(2 * batch_size, max_completion_length)`.
                - `"pixel_values"` (optional): Concatenated pixel values if `"prompt_pixel_values"` are present.
                - `"pixel_attention_mask"` (optional): Concatenated pixel attention masks if `"prompt_pixel_attention_mask"` are present.

        Notes:
            The completion input IDs and attention masks are padded to the maximum completion length of the chosen
            or rejected sequences.
        """
        output = {}

        # For the prompt, the input_ids are the same for both the chosen and rejected responses
        output["prompt_input_ids"] = torch.cat([batch["prompt_input_ids"], batch["prompt_input_ids"]], dim=0)
        output["prompt_attention_mask"] = torch.cat(
            [batch["prompt_attention_mask"], batch["prompt_attention_mask"]], dim=0
        )
        if "pixel_values" in batch:
            output["pixel_values"] = torch.cat([batch["pixel_values"], batch["pixel_values"]], dim=0)

        if "pixel_attention_mask" in batch:
            output["pixel_attention_mask"] = torch.cat(
                [batch["pixel_attention_mask"], batch["pixel_attention_mask"]], dim=0
            )
        if "image_sizes" in batch:
            output["image_sizes"] = torch.cat([batch["image_sizes"], batch["image_sizes"]], dim=0)

        # Concatenate the chosen and rejected completions
        max_completion_length = max(batch["chosen_input_ids"].shape[1], batch["rejected_input_ids"].shape[1])
        output["completion_input_ids"] = torch.cat(
            (
                pad_to_length(batch["chosen_input_ids"], max_completion_length, pad_value=padding_value),
                pad_to_length(batch["rejected_input_ids"], max_completion_length, pad_value=padding_value),
            ),
        )
        output["completion_attention_mask"] = torch.cat(
            (
                pad_to_length(batch["chosen_attention_mask"], max_completion_length, pad_value=0),
                pad_to_length(batch["rejected_attention_mask"], max_completion_length, pad_value=0),
            ),
        )

        return output

    def dpo_loss(
        self,
        chosen_logps: torch.FloatTensor,
        rejected_logps: torch.FloatTensor,
        ref_chosen_logps: torch.FloatTensor,
        ref_rejected_logps: torch.FloatTensor,
    ) -> tuple[torch.FloatTensor, torch.FloatTensor, torch.FloatTensor]:
        """
        Compute the DPO loss for a batch of policy and reference model log probabilities.

        Args:
            chosen_logps (`torch.FloatTensor`):
                Log probabilities of the model for the chosen responses. Shape: `(batch_size,)`.
            rejected_logps (`torch.FloatTensor`):
                Log probabilities of the model for the rejected responses. Shape: `(batch_size,)`.
            ref_chosen_logps (`torch.FloatTensor`):
                Log probabilities of the reference model for the chosen responses. Shape: `(batch_size,)`.
            ref_rejected_logps (`torch.FloatTensor`):
                Log probabilities of the reference model for the rejected responses. Shape: `(batch_size,)`.

        Returns:
            A tuple of three tensors: `(losses, chosen_rewards, rejected_rewards)`.
            The losses tensor contains the DPO loss for each example in the batch.
            The `chosen_rewards` and `rejected_rewards` tensors contain the rewards for the chosen and rejected
            responses, respectively.
        """
        device = self.accelerator.device

        # Get the log ratios for the chosen and rejected responses
        chosen_logratios = chosen_logps.to(device) - (not self.reference_free) * ref_chosen_logps.to(device)
        rejected_logratios = rejected_logps.to(device) - (not self.reference_free) * ref_rejected_logps.to(device)

        if self.f_divergence_type == FDivergenceType.ALPHA_DIVERGENCE.value:
            # The alpha-divergence formula: (1 - u^-alpha) / alpha
            # The divergence difference between the chosen and rejected sample is:
            #     (1 - u[w]^-alpha) / alpha - (1 - u[l]^-alpha) / alpha
            #        = (u[l]^-alpha - u[w]^-alpha) / alpha
            # where u[w] and u[l] are the policy/reference probability ratios
            # for the chosen and rejected samples, respectively.
            alpha_coef = FDivergenceConstants.ALPHA_DIVERGENCE_COEF_DEFAULT
            if self.f_divergence_params and FDivergenceConstants.ALPHA_DIVERGENCE_COEF_KEY in self.f_divergence_params:
                alpha_coef = float(self.f_divergence_params[FDivergenceConstants.ALPHA_DIVERGENCE_COEF_KEY])
            logits = (cap_exp(rejected_logratios * -alpha_coef) - cap_exp(chosen_logratios * -alpha_coef)) / alpha_coef
        else:
            logratios = chosen_logps - rejected_logps
            if self.reference_free:
                ref_logratios = torch.tensor([0], dtype=logratios.dtype, device=logratios.device)
            else:
                ref_logratios = ref_chosen_logps - ref_rejected_logps

            logratios = logratios.to(self.accelerator.device)
            ref_logratios = ref_logratios.to(self.accelerator.device)
            logits = logratios - ref_logratios

            if self.f_divergence_type == FDivergenceType.JS_DIVERGENCE.value:
                # The js-divergence formula: log(2 * u / (1 + u))
                # The divergence difference between the chosen and rejected sample is:
                #     log(2 * u[w] / (1 + u[w])) - log(2 * u[l] / (1 + u[l]))
                #       = log(u[w]) - log(u[l]) - (log(1 + u[w]) - log(1 + u[l]))
                # where u[w] and u[l] are the policy/reference probability ratios
                # for the chosen and rejected samples, respectively.
                logits -= F.softplus(chosen_logratios) - F.softplus(rejected_logratios)

        # The beta is a temperature parameter for the DPO loss, typically something in the range of 0.1 to 0.5.
        # We ignore the reference model as beta -> 0. The label_smoothing parameter encodes our uncertainty about the
        # labels and calculates a conservative DPO loss.
        if self.loss_type == "sigmoid":
            losses = (
                -F.logsigmoid(self.beta * logits) * (1 - self.label_smoothing)
                - F.logsigmoid(-self.beta * logits) * self.label_smoothing
            )

        elif self.loss_type == "robust":
            losses = (
                -F.logsigmoid(self.beta * logits) * (1 - self.label_smoothing)
                + F.logsigmoid(-self.beta * logits) * self.label_smoothing
            ) / (1 - 2 * self.label_smoothing)

        elif self.loss_type == "exo_pair":
            # eqn (16) of the EXO paper: https://huggingface.co/papers/2402.00856
            import math

            if self.label_smoothing == 0:
                self.label_smoothing = 1e-3
            losses = (self.beta * logits).sigmoid() * (
                F.logsigmoid(self.beta * logits) - math.log(1 - self.label_smoothing)
            ) + (-self.beta * logits).sigmoid() * (F.logsigmoid(-self.beta * logits) - math.log(self.label_smoothing))

        elif self.loss_type == "hinge":
            losses = torch.relu(1 - self.beta * logits)

        elif self.loss_type == "ipo":
            # eqn (17) of the paper where beta is the regularization parameter for the IPO loss, denoted by tau in the paper.
            losses = (logits - 1 / (2 * self.beta)) ** 2

        elif self.loss_type == "bco_pair":
            chosen_logratios = chosen_logps - ref_chosen_logps
            rejected_logratios = rejected_logps - ref_rejected_logps
            chosen_rewards = self.beta * chosen_logratios
            rejected_rewards = self.beta * rejected_logratios
            rewards = torch.cat((chosen_rewards, rejected_rewards), 0).mean().detach()
            self.running.update(rewards)
            delta = self.running.mean
            losses = -F.logsigmoid((self.beta * chosen_logratios) - delta) - F.logsigmoid(
                -(self.beta * rejected_logratios - delta)
            )

        elif self.loss_type == "sppo_hard":
            # In the paper (https://huggingface.co/papers/2405.00675), SPPO employs a soft probability approach,
            # estimated using the PairRM score. The probability calculation is conducted outside of the trainer class.
            # The version described here is the hard probability version, where P in Equation (4.7) of Algorithm 1 is
            # set to 1 for the winner and 0 for the loser.
            a = chosen_logps - ref_chosen_logps
            b = rejected_logps - ref_rejected_logps
            losses = (a - 0.5 / self.beta) ** 2 + (b + 0.5 / self.beta) ** 2

        elif self.loss_type == "nca_pair":
            chosen_rewards = (chosen_logps - ref_chosen_logps) * self.beta
            rejected_rewards = (rejected_logps - ref_rejected_logps) * self.beta
            losses = (
                -F.logsigmoid(chosen_rewards)
                - 0.5 * F.logsigmoid(-chosen_rewards)
                - 0.5 * F.logsigmoid(-rejected_rewards)
            )

        elif self.loss_type == "aot_pair":
            chosen_logratios = chosen_logps - ref_chosen_logps
            rejected_logratios = rejected_logps - ref_rejected_logps
            chosen_logratios_sorted, _ = torch.sort(chosen_logratios, dim=0)
            rejected_logratios_sorted, _ = torch.sort(rejected_logratios, dim=0)
            delta = chosen_logratios_sorted - rejected_logratios_sorted
            losses = (
                -F.logsigmoid(self.beta * delta) * (1 - self.label_smoothing)
                - F.logsigmoid(-self.beta * delta) * self.label_smoothing
            )

        elif self.loss_type == "aot":
            logratios = chosen_logps - rejected_logps
            ref_logratios = ref_chosen_logps - ref_rejected_logps
            logratios_sorted, _ = torch.sort(logratios, dim=0)
            ref_logratios_sorted, _ = torch.sort(ref_logratios, dim=0)
            delta = logratios_sorted - ref_logratios_sorted
            losses = (
                -F.logsigmoid(self.beta * delta) * (1 - self.label_smoothing)
                - F.logsigmoid(-self.beta * delta) * self.label_smoothing
            )

        elif self.loss_type == "apo_zero":
            # Eqn (7) of the APO paper (https://huggingface.co/papers/2408.06266)
            # Use this loss when you believe the chosen outputs are better than your model's default output
            losses_chosen = 1 - F.sigmoid(self.beta * chosen_logratios)  # Increase chosen likelihood
            losses_rejected = F.sigmoid(self.beta * rejected_logratios)  # Decrease rejected likelihood
            losses = losses_chosen + losses_rejected

        elif self.loss_type == "apo_down":
            # Eqn (8) of the APO paper (https://huggingface.co/papers/2408.06266)
            # Use this loss when you believe the chosen outputs are worse than your model's default output.
            # Decrease chosen likelihood and decrease rejected likelihood more
            losses_chosen = F.sigmoid(self.beta * chosen_logratios)
            losses_rejected = 1 - F.sigmoid(self.beta * (chosen_logratios - rejected_logratios))
            losses = losses_chosen + losses_rejected

        elif self.loss_type == "discopop":
            # Eqn (5) of the DiscoPOP paper (https://huggingface.co/papers/2406.08414)
            # This loss was discovered with LLM discovery
            logratios = chosen_logps - rejected_logps
            ref_logratios = ref_chosen_logps - ref_rejected_logps
            logits = logratios - ref_logratios
            logits = logits * self.beta
            # Modulate the mixing coefficient based on the log ratio magnitudes
            log_ratio_modulation = torch.sigmoid(logits / self.args.discopop_tau)
            logistic_component = -F.logsigmoid(logits)
            exp_component = torch.exp(-logits)
            # Blend between logistic and exponential component based on log ratio modulation
            losses = logistic_component * (1 - log_ratio_modulation) + exp_component * log_ratio_modulation

        else:
            raise ValueError(
                f"Unknown loss type: {self.loss_type}. Should be one of ['sigmoid', 'hinge', 'ipo', 'exo_pair', "
                "'nca_pair', 'robust', 'bco_pair', 'sppo_hard', 'aot', 'aot_pair', 'discopop', 'apo_zero', 'apo_down']"
            )

        chosen_rewards = self.beta * (chosen_logps.to(device) - ref_chosen_logps.to(device)).detach()
        rejected_rewards = self.beta * (rejected_logps.to(device) - ref_rejected_logps.to(device)).detach()

        return losses, chosen_rewards, rejected_rewards

    def concatenated_forward(self, model: nn.Module, batch: dict[str, Union[list, torch.LongTensor]]):
        """Run the given model on the given batch of inputs, concatenating the chosen and rejected inputs together.

        We do this to avoid doing two forward passes, because it's faster for FSDP.
        """
        num_examples = batch["prompt_input_ids"].shape[0]

        concatenated_batch = self.concatenated_inputs(batch, padding_value=self.padding_value)

        model_kwargs = {}
        if self.aux_loss_enabled:
            model_kwargs["output_router_logits"] = True

        # Add the pixel values and attention masks for vision models
        if "pixel_values" in concatenated_batch:
            model_kwargs["pixel_values"] = concatenated_batch["pixel_values"]
        if "pixel_attention_mask" in concatenated_batch:
            model_kwargs["pixel_attention_mask"] = concatenated_batch["pixel_attention_mask"]
        if "image_sizes" in concatenated_batch:
            model_kwargs["image_sizes"] = concatenated_batch["image_sizes"]

        prompt_input_ids = concatenated_batch["prompt_input_ids"]
        prompt_attention_mask = concatenated_batch["prompt_attention_mask"]
        completion_input_ids = concatenated_batch["completion_input_ids"]
        completion_attention_mask = concatenated_batch["completion_attention_mask"]
        if self.is_encoder_decoder:
            labels = completion_input_ids
            labels[completion_attention_mask == 0] = self.label_pad_token_id
            outputs = model(
                input_ids=prompt_input_ids,
                attention_mask=prompt_attention_mask,
                labels=labels,  # we need the labels for the logits to be returned
                **model_kwargs,
            )
            logits = outputs.logits
            loss_mask = completion_attention_mask.bool()
        else:
            # Concatenate the prompt and completion inputs
            input_ids = torch.cat((prompt_input_ids, completion_input_ids), dim=1)
            attention_mask = torch.cat((prompt_attention_mask, completion_attention_mask), dim=1)
            # Mask the prompt but not the completion for the loss
            loss_mask = torch.cat(
                (torch.zeros_like(prompt_attention_mask), completion_attention_mask),
                dim=1,
            )

            # Flush left to reduce the memory usage
            # [[0, 0, x, x, x, x],  ->  [[x, x, x, x],
            #  [0, x, x, x, 0, 0]]       [x, x, x, 0]]
            attention_mask, input_ids, loss_mask = flush_left(attention_mask, input_ids, loss_mask)

            # Truncate right
            if self.max_length is not None:
                if self.truncation_mode == "keep_end":
                    input_ids = input_ids[:, -self.max_length :]
                    attention_mask = attention_mask[:, -self.max_length :]
                    loss_mask = loss_mask[:, -self.max_length :]
                elif self.truncation_mode == "keep_start":
                    input_ids = input_ids[:, : self.max_length]
                    attention_mask = attention_mask[:, : self.max_length]
                    loss_mask = loss_mask[:, : self.max_length]
                else:
                    raise ValueError(
                        f"Unknown truncation mode: '{self.truncation_mode}'. Should be one of ['keep_end', "
                        "'keep_start']."
                    )

            if self.use_logits_to_keep:
                # Compute logits_to_keep based on loss_mask pattern:
                # [[0, 0, 0, x, x, x, x],
                #  [0, 0, 0, x, x, x, 0]]
                #         ^ start computing logits from here ([:, -(7-3+1):])
                first_compute_index = loss_mask.nonzero(as_tuple=True)[1].min()
                logits_to_keep = (loss_mask.shape[1] - first_compute_index).item() + 1  # +1 for the first label
                model_kwargs["logits_to_keep"] = logits_to_keep

            if self.padding_free:
                # Flatten the input_ids, position_ids, and loss_mask
                # input_ids = [[a, b, c, 0], ->     input_ids = [[a, b, c, d, e, f, g]]
                #              [d, e, f, g]]     position_ids = [[0, 1, 2, 0, 1, 2, 3]]
                input_ids = input_ids[attention_mask.bool()].unsqueeze(0)
                loss_mask = loss_mask[attention_mask.bool()].unsqueeze(0)
                position_ids = attention_mask.cumsum(1)[attention_mask.bool()].unsqueeze(0) - 1
                model_kwargs["position_ids"] = position_ids
            else:
                model_kwargs["attention_mask"] = attention_mask

            outputs = model(input_ids, **model_kwargs)
            logits = outputs.logits

            # Offset the logits by one to align with the labels
            labels = torch.roll(input_ids, shifts=-1, dims=1)
            loss_mask = torch.roll(loss_mask, shifts=-1, dims=1).bool()

            if self.use_logits_to_keep:
                # Align labels with logits
                # logits:    -,  -, [x2, x3, x4, x5, x6]
                #                     ^ --------- ^       after logits[:, :-1, :]
                # labels:   [y0, y1, y2, y3, y4, y5, y6]
                #                         ^ --------- ^   with logits_to_keep=4, [:, -4:]
                # loss_mask: [0,  0,  0,  1,  1,  1,  1]
                labels = labels[:, -logits_to_keep:]
                loss_mask = loss_mask[:, -logits_to_keep:]

        if logits.shape[:2] != labels.shape[:2]:
            # for llava, the returned logits include the image tokens (placed before the text tokens)
            seq_len = labels.shape[1]
            logits = logits[:, -seq_len:]

        # Compute the log probabilities of the labels
        labels[~loss_mask] = 0  # dummy token; we'll ignore the losses on these tokens later
        per_token_logps = selective_log_softmax(logits, labels)
        per_token_logps[~loss_mask] = 0
        per_token_logps = torch.roll(per_token_logps, shifts=1, dims=1)

        if self.padding_free:
            # Unflatten the per_token_logps (shape: [1, sum_seq_len] -> [batch_size, seq_len])
            batch_size, seq_len = attention_mask.shape
            per_token_logps_ = torch.zeros(
                batch_size, seq_len, device=outputs.logits.device, dtype=outputs.logits.dtype
            )
            per_token_logps_[attention_mask.bool()] = per_token_logps
            per_token_logps = per_token_logps_

        all_logps = per_token_logps.sum(-1)

        output = {}

        if self.use_weighting:
            with torch.no_grad():
                # Eq (2) of the WPO paper: https://huggingface.co/papers/2406.11827
                logprobs = F.log_softmax(logits, dim=-1)
                weights_adjustment_factor = torch.logsumexp(2 * logprobs, dim=-1)  # same as sum(probs**2) in log space
                per_token_logps_adjusted = per_token_logps - weights_adjustment_factor
                all_weights = (per_token_logps_adjusted * loss_mask).sum(-1) / loss_mask.sum(-1)
                chosen_weights = all_weights[:num_examples]
                rejected_weights = all_weights[num_examples:]
                output["policy_weights"] = torch.clamp(torch.exp(chosen_weights + rejected_weights), max=1)

        if self.args.rpo_alpha is not None:
            # Only use the chosen logits for the RPO loss
            chosen_logits = logits[:num_examples]
            chosen_labels = labels[:num_examples]

            # Compute the log probabilities of the labels
            output["nll_loss"] = F.cross_entropy(
                torch.flatten(chosen_logits, end_dim=1), torch.flatten(chosen_labels, end_dim=1), ignore_index=0
            )

        if self.loss_type == "ipo":
            all_logps = all_logps / loss_mask.sum(-1)

        output["chosen_logps"] = all_logps[:num_examples]
        output["rejected_logps"] = all_logps[num_examples:]

        # Compute the mean logits
        if self.padding_free:
            # position_ids contains a sequence of range identifiers (e.g., [[0, 1, 2, 0, 1, 2, 3, ...]]).
            # There are 2*num_examples ranges in total: the first half corresponds to the chosen tokens,
            # and the second half to the rejected tokens.
            # To find the start of the rejected tokens, we look for the num_examples+1-th zero in pos_id.
            split_idx = (position_ids == 0).nonzero(as_tuple=True)[1][num_examples]
            mean_chosen_logits = logits[0, :split_idx][loss_mask[0, :split_idx]].mean()
            mean_rejected_logits = logits[0, split_idx:][loss_mask[0, split_idx:]].mean()
        else:
            mean_chosen_logits = logits[:num_examples][loss_mask[:num_examples]].mean()
            mean_rejected_logits = logits[num_examples:][loss_mask[num_examples:]].mean()

        output["mean_chosen_logits"] = mean_chosen_logits
        output["mean_rejected_logits"] = mean_rejected_logits

        if self.aux_loss_enabled:
            output["aux_loss"] = outputs.aux_loss

        return output

    def get_batch_loss_metrics(
        self,
        model,
        batch: dict[str, Union[list, torch.LongTensor]],
        train_eval: Literal["train", "eval"] = "train",
    ):
        """Compute the DPO loss and other metrics for the given batch of inputs for train or test."""
        metrics = {}

        model_output = self.concatenated_forward(model, batch)

        # if ref_chosen_logps and ref_rejected_logps in batch use them, otherwise use the reference model
        if "ref_chosen_logps" in batch and "ref_rejected_logps" in batch:
            ref_chosen_logps = batch["ref_chosen_logps"]
            ref_rejected_logps = batch["ref_rejected_logps"]
        else:
            ref_chosen_logps, ref_rejected_logps = self.compute_ref_log_probs(batch)

        losses, chosen_rewards, rejected_rewards = self.dpo_loss(
            model_output["chosen_logps"], model_output["rejected_logps"], ref_chosen_logps, ref_rejected_logps
        )
        reward_accuracies = (chosen_rewards > rejected_rewards).float()

        if self.args.rpo_alpha is not None:
            losses = losses + self.args.rpo_alpha * model_output["nll_loss"]  # RPO loss from V3 of the paper

        if self.use_weighting:
            losses = losses * model_output["policy_weights"]

        if self.aux_loss_enabled:
            losses = losses + self.aux_loss_coef * model_output["aux_loss"]

        prefix = "eval_" if train_eval == "eval" else ""
        metrics[f"{prefix}rewards/chosen"] = self.accelerator.gather_for_metrics(chosen_rewards).mean().item()
        metrics[f"{prefix}rewards/rejected"] = self.accelerator.gather_for_metrics(rejected_rewards).mean().item()
        metrics[f"{prefix}rewards/accuracies"] = self.accelerator.gather_for_metrics(reward_accuracies).mean().item()
        metrics[f"{prefix}rewards/margins"] = (
            self.accelerator.gather_for_metrics(chosen_rewards - rejected_rewards).mean().item()
        )
        metrics[f"{prefix}logps/chosen"] = (
            self.accelerator.gather_for_metrics(model_output["chosen_logps"]).detach().mean().item()
        )
        metrics[f"{prefix}logps/rejected"] = (
            self.accelerator.gather_for_metrics(model_output["rejected_logps"]).detach().mean().item()
        )
        metrics[f"{prefix}logits/chosen"] = (
            self.accelerator.gather_for_metrics(model_output["mean_chosen_logits"]).detach().mean().item()
        )
        metrics[f"{prefix}logits/rejected"] = (
            self.accelerator.gather_for_metrics(model_output["mean_rejected_logits"]).detach().mean().item()
        )
        if self.args.rpo_alpha is not None:
            metrics[f"{prefix}nll_loss"] = (
                self.accelerator.gather_for_metrics(model_output["nll_loss"]).detach().mean().item()
            )
        if self.aux_loss_enabled:
            metrics[f"{prefix}aux_loss"] = (
                self.accelerator.gather_for_metrics(model_output["aux_loss"]).detach().mean().item()
            )

        return losses.mean(), metrics

    def compute_loss(
        self,
        model: Union[PreTrainedModel, nn.Module],
        inputs: dict[str, Union[torch.Tensor, Any]],
        return_outputs=False,
        num_items_in_batch=None,
    ) -> Union[torch.Tensor, tuple[torch.Tensor, dict[str, torch.Tensor]]]:
        device_type = "xpu" if is_torch_xpu_available() else "cuda"
        compute_loss_context_manager = (
            amp.autocast(device_type) if self._peft_has_been_casted_to_bf16 else nullcontext()
        )
        with compute_loss_context_manager:
            loss, metrics = self.get_batch_loss_metrics(model, inputs, train_eval="train")

        # Make sure to move the loss to the device the original accumulating loss is at back in the `Trainer` class:
        loss = loss.to(self.args.device)
        # force log the metrics
        self.store_metrics(metrics, train_eval="train")

        if return_outputs:
            return loss, metrics

        return loss

    def generate_from_model_and_ref(self, model, batch: dict[str, torch.LongTensor]) -> tuple[str, str]:
        """Generate samples from the model and reference model for the given batch of inputs."""

        # If one uses `generate_during_eval` with peft + bf16, we need to explicitly call generate with
        # the torch amp context manager as some hidden states are silently casted to full precision.
        device_type = "xpu" if is_torch_xpu_available() else "cuda"
        generate_context_manager = amp.autocast(device_type) if self._peft_has_been_casted_to_bf16 else nullcontext()

        with generate_context_manager:
            policy_output = model.generate(
                input_ids=batch["prompt_input_ids"],
                attention_mask=batch["prompt_attention_mask"],
                max_length=self.max_length,
                do_sample=True,
                pad_token_id=self.padding_value,
            )

            # if ref_output in batch use that otherwise use the reference model
            if "ref_output" in batch:
                ref_output = batch["ref_output"]
            else:
                if self.ref_model is None:
                    with self.null_ref_context():
                        ref_output = self.model.generate(
                            input_ids=batch["prompt_input_ids"],
                            attention_mask=batch["prompt_attention_mask"],
                            max_length=self.max_length,
                            do_sample=True,
                            pad_token_id=self.padding_value,
                        )
                else:
                    ref_output = self.ref_model.generate(
                        input_ids=batch["prompt_input_ids"],
                        attention_mask=batch["prompt_attention_mask"],
                        max_length=self.max_length,
                        do_sample=True,
                        pad_token_id=self.padding_value,
                    )

        policy_output = pad_to_length(policy_output, self.max_length, self.padding_value)
        policy_output_decoded = self.processing_class.batch_decode(policy_output, skip_special_tokens=True)

        ref_output = pad_to_length(ref_output, self.max_length, self.padding_value)
        ref_output_decoded = self.processing_class.batch_decode(ref_output, skip_special_tokens=True)

        return policy_output_decoded, ref_output_decoded

    def prediction_step(
        self,
        model: Union[PreTrainedModel, nn.Module],
        inputs: dict[str, Union[torch.Tensor, Any]],
        prediction_loss_only: bool,
        ignore_keys: Optional[list[str]] = None,
    ):
        if ignore_keys is None:
            if hasattr(model, "config"):
                ignore_keys = getattr(model.config, "keys_to_ignore_at_inference", [])
            else:
                ignore_keys = []

        device_type = "xpu" if is_torch_xpu_available() else "cuda"
        prediction_context_manager = amp.autocast(device_type) if self._peft_has_been_casted_to_bf16 else nullcontext()

        with torch.no_grad(), prediction_context_manager:
            loss, metrics = self.get_batch_loss_metrics(model, inputs, train_eval="eval")

        # force log the metrics
        self.store_metrics(metrics, train_eval="eval")

        if prediction_loss_only:
            return loss.detach(), None, None

        # logits for the chosen and rejected samples from model
        logits_dict = {
            "eval_logits/chosen": metrics["eval_logits/chosen"],
            "eval_logits/rejected": metrics["eval_logits/rejected"],
        }
        logits = tuple(v.unsqueeze(dim=0) for k, v in logits_dict.items() if k not in ignore_keys)
        logits = torch.stack(logits).mean(axis=1).to(self.accelerator.device)
        labels = torch.zeros(logits.shape[0], device=self.accelerator.device)

        return (loss.detach(), logits, labels)

    def store_metrics(self, metrics: dict[str, float], train_eval: Literal["train", "eval"] = "train") -> None:
        for key, value in metrics.items():
            self._stored_metrics[train_eval][key].append(value)

    def evaluation_loop(
        self,
        dataloader: DataLoader,
        description: str,
        prediction_loss_only: Optional[bool] = None,
        ignore_keys: Optional[list[str]] = None,
        metric_key_prefix: str = "eval",
    ) -> EvalLoopOutput:
        """
        Overriding built-in evaluation loop to store metrics for each batch.
        Prediction/evaluation loop, shared by `Trainer.evaluate()` and `Trainer.predict()`.

        Works both with or without labels.
        """

        # Sample and save to game log if requested (for one batch to save time)
        if self.generate_during_eval:
            # Generate random indices within the range of the total number of samples
            num_samples = len(dataloader.dataset)
            random_indices = random.sample(range(num_samples), k=self.args.eval_batch_size)

            # Use dataloader.dataset.select to get the random batch without iterating over the DataLoader
            random_batch_dataset = dataloader.dataset.select(random_indices)
            random_batch = self.data_collator(random_batch_dataset)
            random_batch = self._prepare_inputs(random_batch)

            policy_output_decoded, ref_output_decoded = self.generate_from_model_and_ref(self.model, random_batch)

            table = pd.DataFrame(
                columns=["Prompt", "Policy", "Ref Model"],
                data=[
                    [prompt, pol[len(prompt) :], ref[len(prompt) :]]
                    for prompt, pol, ref in zip(
                        random_batch_dataset["prompt"], policy_output_decoded, ref_output_decoded
                    )
                ],
            )
            if "wandb" in self.args.report_to:
                wandb.log({"game_log": wandb.Table(data=table)})

            if "comet_ml" in self.args.report_to:
                log_table_to_comet_experiment(
                    name="game_log.csv",
                    table=table,
                )

        # Base evaluation
        initial_output = super().evaluation_loop(
            dataloader, description, prediction_loss_only, ignore_keys, metric_key_prefix
        )

        return initial_output

    def log(self, logs: dict[str, float], start_time: Optional[float] = None) -> None:
        """
        Log `logs` on the various objects watching training, including stored metrics.

        Args:
            logs (`dict[str, float]`):
                The values to log.
            start_time (`float` or `None`, *optional*, defaults to `None`):
                Start time of the training.
        """
        # logs either has 'loss' or 'eval_loss'
        train_eval = "train" if "loss" in logs else "eval"
        # Add averaged stored metrics to logs
        for key, metrics in self._stored_metrics[train_eval].items():
            logs[key] = torch.tensor(metrics).mean().item()
        del self._stored_metrics[train_eval]

        if version.parse(transformers.__version__) >= version.parse("4.47.0.dev0"):
            return super().log(logs, start_time)
        else:  # transformers<=4.46
            return super().log(logs)

    def create_model_card(
        self,
        model_name: Optional[str] = None,
        dataset_name: Optional[str] = None,
        tags: Union[str, list[str], None] = None,
    ):
        """
        Creates a draft of a model card using the information available to the `Trainer`.

        Args:
            model_name (`str` or `None`, *optional*, defaults to `None`):
                Name of the model.
            dataset_name (`str` or `None`, *optional*, defaults to `None`):
                Name of the dataset used for training.
            tags (`str`, `list[str]` or `None`, *optional*, defaults to `None`):
                Tags to be associated with the model card.
        """
        if not self.is_world_process_zero():
            return

        if hasattr(self.model.config, "_name_or_path") and not os.path.isdir(self.model.config._name_or_path):
            base_model = self.model.config._name_or_path
        else:
            base_model = None

        tags = tags or []
        if isinstance(tags, str):
            tags = [tags]

        if hasattr(self.model.config, "unsloth_version"):
            tags.append("unsloth")

        citation = textwrap.dedent(
            """\
            @inproceedings{rafailov2023direct,
                title        = {{Direct Preference Optimization: Your Language Model is Secretly a Reward Model}},
                author       = {Rafael Rafailov and Archit Sharma and Eric Mitchell and Christopher D. Manning and Stefano Ermon and Chelsea Finn},
                year         = 2023,
                booktitle    = {Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023},
                url          = {http://papers.nips.cc/paper_files/paper/2023/hash/a85b405ed65c6477a4fe8302b5e06ce7-Abstract-Conference.html},
                editor       = {Alice Oh and Tristan Naumann and Amir Globerson and Kate Saenko and Moritz Hardt and Sergey Levine},
            }"""
        )

        model_card = generate_model_card(
            base_model=base_model,
            model_name=model_name,
            hub_model_id=self.hub_model_id,
            dataset_name=dataset_name,
            tags=tags,
            wandb_url=wandb.run.get_url() if is_wandb_available() and wandb.run is not None else None,
            comet_url=get_comet_experiment_url(),
            trainer_name="DPO",
            trainer_citation=citation,
            paper_title="Direct Preference Optimization: Your Language Model is Secretly a Reward Model",
            paper_id="2305.18290",
        )

        model_card.save(os.path.join(self.args.output_dir, "README.md"))
class UnslothDPOTrainer(_UnslothDPOTrainer):
    """
    
    Initialize DPOTrainer.

    Args:
        model (`transformers.PreTrainedModel`):
            The model to train, preferably an `AutoModelForSequenceClassification`.
        ref_model (`PreTrainedModelWrapper`):
            Hugging Face transformer model with a casual language modelling head. Used for implicit reward computation and loss. If no
            reference model is provided, the trainer will create a reference model with the same architecture as the model to be optimized.
        args (`DPOConfig`):
            The DPO config arguments to use for training.
        data_collator (`transformers.DataCollator`):
            The data collator to use for training. If None is specified, the default data collator (`DataCollatorForPreference`) will be used
            which will pad the sequences to the maximum length of the sequences in the batch, given a dataset of paired sequences.
        train_dataset (`datasets.Dataset`):
            The dataset to use for training.
        eval_dataset (`datasets.Dataset`):
            The dataset to use for evaluation.
        processing_class (`PreTrainedTokenizerBase` or `BaseImageProcessor` or `FeatureExtractionMixin` or `ProcessorMixin`, *optional*):
            Processing class used to process the data. If provided, will be used to automatically process the inputs
            for the model, and it will be saved along the model to make it easier to rerun an interrupted training or
            reuse the fine-tuned model.
            This supercedes the `tokenizer` argument, which is now deprecated.
        model_init (`Callable[[], transformers.PreTrainedModel]`):
            The model initializer to use for training. If None is specified, the default model initializer will be used.
        compute_metrics (`Callable[[EvalPrediction], dict]`, *optional*):
            The function to use to compute the metrics. Must take a `EvalPrediction` and return
            a dictionary string to metric values.
        callbacks (`list[transformers.TrainerCallback]`):
            The callbacks to use for training.
        optimizers (`tuple[torch.optim.Optimizer, torch.optim.lr_scheduler.LambdaLR]`):
            The optimizer and scheduler to use for training.
        preprocess_logits_for_metrics (`Callable[[torch.Tensor, torch.Tensor], torch.Tensor]`):
            The function to use to preprocess the logits before computing the metrics.
        peft_config (`dict`, defaults to `None`):
            The PEFT configuration to use for training. If you pass a PEFT configuration, the model will be wrapped in a PEFT model.
    
    """
    def __init__(
        self,
        model = None,
        ref_model = None,
        args = None,
        data_collator = None,
        train_dataset = None,
        eval_dataset = None,
        processing_class = None,
        model_init = None,
        compute_metrics = None,
        callbacks = None,
        preprocess_logits_for_metrics = None,
        peft_config = None,
        **kwargs
    ):
        if args is None: args = UnslothDPOConfig()
        use_bf16 = getattr(args, 'bf16', False)
        use_fp16 = getattr(args, 'fp16', False)
        force_float32 = False
        if os.environ.get('UNSLOTH_FORCE_FLOAT32', '0') == '1':
            print('Unsloth: Switching to float32 training since model cannot work with float16')
            force_float32 = True
        mixed_precision_dtype = os.environ.get('UNSLOTH_MIXED_PRECISION', 'float32')
        dtype = getattr(model.config, 'torch_dtype', None)
        if dtype is None: dtype = model.get_input_embeddings().dtype
        from unsloth_zoo.utils import _get_dtype
        dtype = _get_dtype(dtype)
        float16 = dtype == torch.float16
        if not force_float32 and (float16 and use_bf16): raise TypeError('Unsloth: Model is in float16 precision but you want to use bfloat16 precision. Set fp16 to `True` and bf16 to `False`')
        if not force_float32 and (not float16 and use_fp16): raise TypeError('Unsloth: Model is in bfloat16 precision but you want to use float16 precision. Set fp16 to `False` and bf16 to `True`')
        if force_float32:
            args.fp16 = False
            args.bf16 = False
            os.environ['ACCELERATE_MIXED_PRECISION'] = 'no'
        elif (not use_bf16 and not use_fp16) and mixed_precision_dtype == 'float32':
            args.fp16 = float16
            args.bf16 = not float16
            os.environ['ACCELERATE_MIXED_PRECISION'] = 'fp16' if float16 else 'bf16'
        if getattr(args, 'eval_dataset', None) is not None and getattr(args, 'eval_strategy', 'no') == 'no':
            args.eval_strategy = 'steps'
            if getattr(args, 'eval_steps', None) is None: args.eval_steps = 0.1
        ga_steps = getattr(args, 'gradient_accumulation_steps', None)
        if ga_steps is not None and ga_steps > 1:
            from transformers import __version__ as transformers_version
            if Version(transformers_version) <= Version('4.45.2'):
                print('**** Unsloth: Please use our fixed gradient_accumulation_steps by updating transformers, TRL and Unsloth!\n'
                      '`pip install --upgrade --no-cache-dir --force-reinstall --no-deps unsloth transformers trl unsloth_zoo`')
        if getattr(args, 'eval_strategy', 'no') != 'no':
            eval_bsz = getattr(args, 'per_device_eval_batch_size', 8)
            if eval_bsz == 8 and args.per_device_train_batch_size < eval_bsz: args.per_device_eval_batch_size = args.per_device_train_batch_size
            if getattr(args, 'eval_accumulation_steps', None) is None and ga_steps is not None: args.eval_accumulation_steps = ga_steps
        fp16_full_eval = getattr(args, 'fp16_full_eval', False)
        bf16_full_eval = getattr(args, 'bf16_full_eval', False)
        if args.fp16 and bf16_full_eval: args.bf16_full_eval = False; args.fp16_full_eval = True
        if args.bf16 and fp16_full_eval: args.bf16_full_eval = True; args.fp16_full_eval = False
        if force_float32:
            args.bf16_full_eval = False
            args.fp16_full_eval = False
        elif os.environ.get('UNSLOTH_MIXED_PRECISION', 'float32') == 'bfloat16':
            args.bf16_full_eval = True
            args.fp16_full_eval = False
        elif not bf16_full_eval and not fp16_full_eval:
            args.bf16_full_eval = args.bf16
            args.fp16_full_eval = args.fp16
        _output_logits = False
        if locals().get('compute_metrics', None) is not None: _output_logits = True
        if locals().get('preprocess_logits_for_metrics', None) is not None: _output_logits = True
        if _output_logits:
            os.environ['UNSLOTH_RETURN_LOGITS'] = '1'
        if 'max_seq_length' not in locals() and not hasattr(args, 'max_seq_length'):
            pass
        else:
            model_max_seq_length = getattr(model, 'max_seq_length', None)
            args_max_seq_length  = getattr(args,  'max_seq_length', None)
            if args_max_seq_length is None and model_max_seq_length is not None:
                max_seq_length = model.max_seq_length
                if hasattr(args, 'max_seq_length'): args.max_seq_length = max_seq_length
        if model is not None and hasattr(model, 'for_training'):
            model.for_training()
        if 'tokenizer' in locals() and hasattr(tokenizer, 'padding_side'): tokenizer.padding_side = 'right'
        if 'processing_class' in locals():
            if hasattr(processing_class, 'padding_side'): processing_class.padding_side = 'right'
            if hasattr(processing_class, 'tokenizer') and hasattr(processing_class.tokenizer, 'padding_side'): processing_class.tokenizer.padding_side = 'right'
        __tokenizer = processing_class if 'processing_class' in locals() else tokenizer
        from unsloth_zoo.vision_utils import UnslothVisionDataCollator
        if not isinstance(data_collator, UnslothVisionDataCollator):
            if isinstance(data_collator, DataCollatorForSeq2Seq) and 'labels' not in train_dataset.column_names:
                data_collator = DataCollatorForLanguageModeling(__tokenizer, mlm = False)
            elif isinstance(data_collator, DataCollatorForLanguageModeling) and 'labels' in train_dataset.column_names:
                data_collator = DataCollatorForSeq2Seq(__tokenizer)
        else:
            if hasattr(args, 'remove_unused_columns'): args.remove_unused_columns = False
            if hasattr(args, 'dataset_text_field'): args.dataset_text_field = ''
            if hasattr(args, 'dataset_kwargs'): args.dataset_kwargs = {'skip_prepare_dataset': True}
        if not isinstance(data_collator, UnslothVisionDataCollator):
            if not hasattr(__tokenizer, 'pad') and hasattr(__tokenizer, 'tokenizer'):
                if isinstance(data_collator, DataCollatorForSeq2Seq):
                    data_collator = DataCollatorForSeq2Seq(__tokenizer.tokenizer)
                else:
                    data_collator = DataCollatorForLanguageModeling(__tokenizer.tokenizer, mlm = False)
        other_metrics = []
        
        from unsloth_zoo.logging_utils import PatchRLStatistics
        PatchRLStatistics('dpo_trainer', other_metrics)
        if hasattr(train_dataset, 'column_names'):
            column_names = set(train_dataset.column_names)
            check = ['chosen', 'rejected', 'prompt', 'chosen_input_ids', 'chosen_attention_mask',
                     'chosen_labels', 'rejected_input_ids', 'rejected_attention_mask', 'rejected_labels',
                     'prompt_input_ids', 'prompt_attention_mask']
            if all(x in column_names for x in check):
                train_dataset = train_dataset.remove_columns(['chosen', 'rejected', 'prompt'])
            del check, column_names
        
        super().__init__(
            model = model,
            ref_model = ref_model,
            args = args,
            data_collator = data_collator,
            train_dataset = train_dataset,
            eval_dataset = eval_dataset,
            processing_class = processing_class,
            model_init = model_init,
            compute_metrics = compute_metrics,
            callbacks = callbacks,
            preprocess_logits_for_metrics = preprocess_logits_for_metrics,
            peft_config = peft_config,**kwargs)
        if hasattr(self, 'neftune_hook_handle'):
            self.neftune_hook_handle.remove()
            if hasattr(self, 'neftune_hook_handle'): del self.neftune_hook_handle
        if getattr(args, 'neftune_noise_alpha', None) is not None:
            model.get_input_embeddings().neftune_noise_alpha = self.neftune_noise_alpha
        pass
        
pass